Research On Human Posture Recognition Algorithm Based On Deep Learning

With the rapid development of computer vision and deep learning technology,human posture recognition has gradually become an important research field.In recent years,human posture recognition technology has been widely used in sports.This also makes the research of human posture recognition technology more meaningful in practical applications.At present,there are three main types of human posture recognition research methods based on deep learning,which are 2D(2-Dimensional)convolutional two-stream networks,3D(3-Dimensional)convolutional neural networks,and cyclic convolutional networks.Among them,3D convolution can directly extract spatiotemporal features without feature fusion.Therefore,the use of this research method is more common.C3D(Convolutional 3D)network is a classic network of 3D convolution.This network has important research significance in the field of human posture recognition based on 3D convolution.Therefore,this paper is based on C3 D network to carry out research on human posture recognition.Since the 3D convolution in the C3 D network increases the time dimension compared with the 2D convolution,the parameter of the C3 D network is relatively large,and the structure of the C3 D network is relatively simple,so the accuracy of human posture recognition needs to be improved.In response to the above problems,this paper designs a new network structure based on C3 D network,which can reduce the number of C3 D network parameters and further improve the performance of human posture recognition.The main research contents are as follows:(1)Human posture recognition based on improved C3 D network: On the basis of C3 D network,global average pooling is used to replace the fully connected layer,and the asymmetric3 D convolutional layer is constructed through the splitting and merging of the convolution kernel.Then,3D point convolutional layers and batch normalization are introduced.Finally,the Re LU activation function was replaced with the GELU activation function,and finally the improved C3 D network was proposed for human posture recognition.(2)Human posture recognition based on C3 D attentional residual network: Based on the improved C3 D network,it is extended to the residual network with full pre-activated residual structure,and use soft pooling to replace the maximum pooling in the network.Then,group normalization is used to replace batch normalization.Finally,the spatiotemporal channel attention mechanism is introduced into the network,so as to propose the C3 D attention full preactivated residual network.Then,the full pre-activated residual structure is replaced with the staged residual structure,and finally,the C3 D attention staged residual network is proposed for human posture recognition.In this paper,the various improved networks are compared with C3 D networks on HMDB51 dataset to verify the effectiveness of the proposed improved method.Then,the C3 D attention staged residual network is compared with other popular networks on the self-built 43 categories of sports data sets,so as to verify that the network in this paper has good applicability in sports recognition.Subsequently,a comparative experiment was carried out on the swimming data set,which verified that the network in this paper also has a good recognition effect in the fine-grained action recognition of sports.